Graph neural networks for spatial-temporal analysis of defect propagation in layer-based manufacturing processes

Layer-based quality control in additive manufacturing is challenging due to the spatio-temporal complexity of defect development and propagation across build layers. Most deep learning approaches perform layer-independent analysis, limiting their ability to predict defect evolution. This work proposes a CNN Spatio-Temporal Graph Neural Network (CNN-STGNN) framework for defect recognition, severity prediction, and early warning in layer-by-layer manufacturing. A one-dimensional CNN first encodes multimodal in-situ signals—including process parameters, thermal features, and acoustic descriptors—into spatial embeddings. These embeddings construct a spatio-temporal graph where nodes represent layer-region pairs and edges capture intra-layer spatial adjacency and inter-layer temporal continuity. A graph attention mechanism identifies critical layers and dominant fault propagation paths. Evaluated on 1,000 samples spanning 100 layers and five spatial regions, CNN-STGNN achieves 85.9% defect classification accuracy and a 0.84 macro F1-score, a 13.1% improvement over CNN-only baselines. Severity estimation yields an RMSE of 0.061 and R² of 0.89. Temporal analysis demonstrates detection of defects up to 25 layers ahead with over 74% accuracy and below 19% false alarm rate. Graph attention visualizations further provide interpretable identification of high-risk layers and propagation pathways, enabling proactive quality control in industrial additive manufacturing.